Crude estimate as of 2022 of highly variable cache size and performance. Depends heavily on purpose of processor (low power laptop vs high performance computing / gaming server)
SWAP Gibibytes milliseconds
DRAM Gibibytes ~50+ nanoseconds
L3 cache ~100 MiB ~20-40 ns
L2 cache ~10 MiB ~5-10 ns
L1 cache ~1 MiB ~1 ns
The less memory a program uses, the more likely most of its data will fit in faster levels of memory, and the lower the average memory access time (AMAT).
Average disk access time for a random block is a few milliseconds, write buffering, disk caching and predictive reads reduce this.
Multiple levels, hit ratio.
Restricting memory access to a smaller range improves hit ratio and can have a dramatic impact on program speed.
Filling a 64.00 KiB array 16384 times 1 bytes at a time...
64.00 KiB array 1.00 B blocks 653.00 ms 1568.15 MiB/s
Filling a 64.00 KiB array 16384 times 2 bytes at a time...
64.00 KiB array 2.00 B blocks 263.00 ms 3893.54 MiB/s
Filling a 64.00 KiB array 16384 times 4 bytes at a time...
64.00 KiB array 4.00 B blocks 147.00 ms 6965.99 MiB/s
Filling a 64.00 KiB array 16384 times 8 bytes at a time...
64.00 KiB array 8.00 B blocks 89.00 ms 11505.62 MiB/s
Testing large array, low cache hit-ratio...
Filling a 512.00 MiB array 5 times 1 bytes at a time...
512.00 MiB array 1.00 B blocks 1905.00 ms 1343.83 MiB/s
Filling a 512.00 MiB array 5 times 2 bytes at a time...
512.00 MiB array 2.00 B blocks 681.00 ms 3759.18 MiB/s
Filling a 512.00 MiB array 5 times 4 bytes at a time...
512.00 MiB array 4.00 B blocks 525.00 ms 4876.19 MiB/s
Filling a 512.00 MiB array 5 times 8 bytes at a time...
512.00 MiB array 8.00 B blocks 509.00 ms 5029.47 MiB/s
Example: 30ns DRAM, 2ns cache. Hit ratio 0.9, avg access time is 0.9 * 2 + 0.1 * 30 = 4.8 ns. Hit ration 0.5, avg access time is 0.5 * 2 + 0.5 * 30 = 16 ns.
With 1% VM swapping: 4ms disk (1ms avg access time), 30ns DRAM, 2ns cache. 0.01 * 1,000,000ns + 0.69 * 30ns + 0.3 * 2ns = 10,021.3ns